A risk entropy self-adaptive adjustment perception degradation security planning method and system

Abstract

The present application relates to a kind of risk entropy self-adapting regulation's perception degradation safety planning method and system, the method includes: M1.the reception and pre-processing of multi-source perception data;M2.perception quality evaluation and perception degradation determination, based on the perception confidence and covariance matrix of step 1 output, carry out three levels of evaluation processing;M3.risk entropy calculation and confidence-expansion coefficient mapping, risk entropy calculation engine is jointly modeled into probability distribution with the perception confidence and predicted covariance matrix, calculate risk entropy value H, and establish the quantitative mapping relationship of risk entropy and safety corridor expansion coefficient.The present application not only solves the problem that perception quality information cannot enter planning optimization solver in the prior art, but also avoids the planning behavior mutation caused by discrete trajectory set switching.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a perception degradation safety planning method and system with adaptive risk entropy adjustment. Background Technology

[0002] Autonomous driving technology has made significant progress in perception, decision-making, and control in recent years. Local path planning, as a key module connecting upper-level behavioral decision-making and lower-level vehicle control, directly determines the driving safety, traffic efficiency, and passenger comfort of autonomous vehicles. In complex and dynamic traffic scenarios, vehicles need to respond in real time to changes in the movement of surrounding traffic participants, generating smooth and executable local trajectories while meeting traffic rules, vehicle dynamics constraints, and collision avoidance requirements. This capability has become one of the core technological challenges for the commercialization of high-level autonomous driving systems (L3 and above).

[0003] In existing technologies, local path planning methods are mainly divided into three categories: search-based planning methods, sampling-based planning methods, and optimization-based planning methods. Search-based methods (such as A* and JPS) can find the shortest path in a known static map, but the computational complexity increases exponentially with the dimension of the search space, making it difficult to meet the real-time requirements of high-speed driving. Sampling-based methods (such as RRT and RRT*) are more efficient in exploring high-dimensional spaces, but the generated paths usually lack smoothness and require additional post-processing steps. Optimization-based methods model path planning as a constrained numerical optimization problem, which can simultaneously consider safety, smoothness, and dynamic feasibility, and is the current mainstream research direction.

[0004] However, traditional optimization-based local path planning methods still face the following technical bottlenecks when dealing with scenarios of perception performance degradation: (1) The safety corridor expansion coefficient is fixed and cannot be dynamically adjusted according to perception confidence, resulting in insufficient safety margin when sensor performance deteriorates; (2) There is a lack of coupling mechanism between perception quality and risk assessment. When perception confidence decreases, the planning result still maintains the same constraint strength as when perception is normal, resulting in insufficient planning safety; (3) The planning strategy lacks the ability to respond hierarchically in scenarios of perception degradation and cannot smoothly switch to conservative planning mode when perception deteriorates; (4) There is a lack of end-to-end closed loop between perception degradation and safety planning decision. Open-loop planning cannot adjust planning behavior in real time when perception fails.

[0005] In the prior art, Chinese patent (publication number: CN122166145A) discloses an autonomous driving behavior control method and system based on perception reliability priors. Addressing the soft anomaly problem of "appearing usable but actually distorted" perception results in autonomous driving systems, this method extracts at least three types of runtime features, including sensor health, multimodal consistency, and temporal stability, and jointly generates local and global perception reliability priors. This system operates independently of the feature fusion process within the perception network, directly performing hierarchical behavior gating at the kinematic envelope level: restricting local driving actions for specific targets or regions based on local priors; and shrinking or performing degraded control on the candidate trajectory set based on global priors. However, this scheme focuses on post-event diagnosis and early warning of perception failure, without establishing a dynamic coupling relationship between perception confidence and planning constraint strength. Early warning information is only transmitted to the decision-making module but does not participate in the real-time adjustment of the planning process. It does not establish a mapping mechanism between perception quality level and planning strategy, cannot adaptively switch to conservative planning mode according to the degree of perception degradation, and the response of planning strategy lags behind the degradation of perception quality. It does not quantify perception confidence into risk parameters (such as risk entropy) that can participate in optimization calculations, cannot explicitly model the impact of perception uncertainty on safety constraints in the planning solver, and failure handling is mainly based on early warning signals and compensatory recovery. It lacks an adaptive adjustment mechanism for the safety corridor expansion coefficient when perception degradation occurs, and cannot expand the safety margin in real time when perception performance degrades.

[0006] In the prior art, Chinese patent (publication number: CN113741388B) discloses a safety diagnosis system and method based on autonomous driving perception failure, including necessary perception sensors and decision-making modules for autonomous driving. The perception sensors include LiDAR, cameras, and millimeter-wave radar. The system also includes a failure mechanism submodule, a failure diagnosis subsystem, and a failure handling subsystem. The failure mechanism submodule diagnoses based on relevant information from the perception sensor circuits to determine the fault propagation mechanism and the specific impact of the fault on perception performance. The failure diagnosis subsystem diagnoses based on data quality and driving algorithms. The perception failure handling subsystem integrates information from the failure mechanism submodule and the failure diagnosis subsystem, performs compensatory recovery and algorithm supplementation, and outputs a warning signal to the decision-making module. However, the candidate trajectory set shrinkage in this scheme is a discrete trajectory selection mechanism (selecting from a preset trajectory cluster) rather than an end-to-end constrained optimization planning method. It lacks explicit modeling and joint solution of vehicle dynamics constraints, traffic rule constraints, and comfort constraints. It does not establish a continuous dynamic mapping relationship between perception confidence and safety corridor expansion coefficient, and cannot achieve fine-grained continuous adjustment of safety margin. The trajectory set shrinkage is a coarse-grained discrete switch rather than a smooth transition. It does not quantify perception uncertainty into risk entropy parameters and introduce them into the cost function of the constrained optimization solver. Perception quality information is only used for behavior gating rather than participating in the construction of optimization objectives. It does not establish a closed-loop mechanism between perception degradation and execution feedback. When the tracking deviation exceeds the threshold, it lacks a dynamic update and replanning trigger strategy for the safety corridor. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the present invention provides a perception degradation safety planning method and system with risk entropy adaptive adjustment, which not only solves the problem that perception quality information cannot enter the planning optimization solver in the prior art, but also avoids the sudden change in planning behavior caused by the switching of discrete trajectory sets.

[0008] To achieve the above and other related objectives, the present invention provides the following technical solution: A risk entropy adaptive adjustment-based perceived degradation safety planning method, the method comprising: M1. Reception and preprocessing of multi-source sensing data; M2. Perception quality assessment and perception degradation determination: Based on the perception confidence and covariance matrix output in step 1, three levels of assessment processing are performed. M3. Risk Entropy Calculation and Confidence-Inflation Coefficient Mapping: The risk entropy calculation engine jointly models the perceived confidence and the predicted covariance matrix as a probability distribution, calculates the risk entropy value H, and establishes a quantitative mapping relationship between risk entropy and the safety corridor inflation coefficient. M4. Adaptive safety corridor dynamic generation: Based on the safety corridor expansion coefficient output in step 3, and combined with the drivable area boundary constraints extracted from the high-precision map, a safety corridor is dynamically generated. M5. Planning strategy hierarchical selection and parameter configuration: Based on the perception degradation level output in step 2, select the current planning strategy from normal mode, early warning mode and conservative mode, and dynamically configure the parameter set of the constraint optimization planner. M6. Constraint optimization trajectory solution: Receive the adaptive safety corridor constraints generated in step 4 and the planning strategy parameters configured in step 5, construct and solve the multi-objective constrained optimization problem; M7. Trajectory Output and Execution Feedback: After the optimization solution converges, the constraint optimization planner outputs the optimal trajectory that satisfies all priority constraints, including the trajectory position sequence, the corresponding velocity sequence, and the acceleration sequence. M8. Perception Degradation Continuous Monitoring and Dynamic Update: Throughout the entire trajectory execution process, the perception quality assessment module continuously executes the assessment process of step 2 to monitor perception quality in real time.

[0009] Furthermore, in step M1, the multi-source sensing data reception and preprocessing involves the input end receiving in real time a target list and corresponding sensing confidence scores from the vehicle-mounted sensing system; dynamic targets such as vehicles and pedestrians perceived by the camera in front and to the side, accompanied by target category confidence scores; target distance, relative speed, and angle information provided by millimeter-wave radar, accompanied by radial speed confidence scores; and high-precision three-dimensional sensing target point clouds generated by lidar, accompanied by point cloud density confidence scores. Simultaneously, the trajectory prediction module generates a multimodal predicted trajectory and its covariance matrix for each sensing target within a future set time domain.

[0010] Furthermore, in step M2, the three-level evaluation process based on the perceived confidence and covariance matrix output in step 1 includes: M21. Sensor-level confidence fusion: For each perceived target i, the sensor-level confidence fusion unit performs exponentially weighted smoothing on the perceived confidence of its historical T frames. , Where α∈(0,1) are the fusion weight coefficients, c i (t) represents the original perceived confidence of the i-th target at the current time, and c fused i The confidence level of the target-level perception after fusion; M22. Scene-level perception quality score: The scene-level perception quality score unit integrates the fusion confidence distribution of all perceived targets, the rate of change in the number of targets, and the detection missing rate to calculate the scene-level perception quality score Q. , , in, r is the mean confidence score of all perceived targets fused together. miss r represents the detection missing rate of the current frame compared to the previous frame. drift The target quantity change rate is represented by w1, w2, and w3, which are weighting coefficients. When Q ≥ Q normal When Q is determined to be in a normal perception state, degraded ≤Q normal When Q is determined to be in a downgraded perception state, degraded The system is judged to be in a severely degraded perception state. degraded and Q normal The preset perceived quality threshold; M23. Perception Degradation Judgment and Level Output: The perception degradation judgment unit outputs the perception degradation level L based on the comparison between the scene-level perception quality score Q and the threshold. .

[0011] Furthermore, the constraint condition for the weighting coefficients w1, w2, and w3 is w1 + w2 + w3 = 1.

[0012] Furthermore, in step M3, the risk entropy calculation and confidence-inflation coefficient mapping includes: M31. Uncertainty joint probability distribution modeling, fusing the confidence level c of each perceived target i. fused i And the predicted covariance matrix ∑ i Joint modeling is performed on the vehicle at each discrete position p. k The probability distribution p at point k (p), , Where, μ k The center position of the k-th discrete risk interval; M32. Risk entropy calculation, based on the probability distribution p in step 31. k (p), calculate the risk entropy value H, which describes the degree of uncertainty in the vehicle's location. , The risk entropy H(p) ranges from [0, logK], where H(p) = 0 indicates that the perception is completely certain, and H(p) = logK indicates that the perception is completely uncertain. M33. Entropy-Confidence Mapping and Inflation Coefficient Calculation: Establishing a quantitative mapping relationship between the risk entropy value H(p) and the safety corridor inflation coefficient δ. When the perceived confidence decreases, the risk entropy increases, the safety corridor inflation coefficient increases accordingly, and the safety margin automatically expands. ​​, Where, δ min λ is the minimum width of the safety corridor. H δ is the risk entropy amplification factor. k Given the current expansion coefficient of the safety corridor, δ increases as H(p) increases. k It increases linearly thereafter; when H(p) = 0, δ k =δ min It degenerates to the basic width under normal perception conditions.

[0013] Furthermore, in step M33, a correction is introduced to the calculation of the expansion coefficient based on the perception degradation level L. , Where, β L This is a correction factor for perception degradation level, used to further increase the safety margin when a severe perception degradation is detected.

[0014] Furthermore, in step M4, using the drivable area boundary as the inner constraint and the expanded risk area boundary as the outer constraint, the non-convex safety area is convexened into a series of convex polygon segments using a sequential convexification method: , Among them, C safe,k For the k-th safe corridor, A k and b k Let p be the constraint matrix and constraint vector of the convex polygon, respectively. ref,k δ is the reference path point for this section of the corridor. k Let C be the expansion coefficient of this corridor segment, and C be the safety corridor constraint. safe It will serve as the hard constraint boundary for the constrained optimization planner.

[0015] Furthermore, in step M5, when the perceived degradation level L=0, the normal mode is selected: all planning parameters remain at their default values, and the safety margin is δ. min Based on this, the traffic efficiency weight w e Maintain a high level; when L=1, select the early warning mode: increase the safety margin to δ. min +λ H • H(p), traffic efficiency weight w e Reduce the safety constraint weight w s Increase, while introducing an additional uncertainty penalty term w u • H(p) is incorporated into the cost function; when L=2, a conservative mode is selected: the safety margin is further increased to δ. min +λ H ∙H(p)+β L • L, traffic efficiency weight w e Reduced to the minimum, safety constraint weight ws Increase to the highest setting, limiting the maximum speed to v. safe max And increase the collision detection safety distance d safe .

[0016] Furthermore, in step M6, considering four dimensions—safety, smoothness, traffic efficiency, and risk margin—the cost function J for the trajectory optimization problem is constructed. , Among them, w u For uncertain penalty weights, J uncertainty For uncertainty cost terms based on risk entropy, , Security cost function J safety (τ) Penalty trajectory for crossing high-risk areas, smoothing cost function J smooth (τ) The jerk of the penalty trajectory is used to ensure ride comfort; the cost function J for traffic efficiency. efficiency (τ) Encourages the trajectory to reach the target point quickly, risk margin cost function J risk,margin (τ) The distance between the penalty trajectory and the high-risk area is insufficient.

[0017] To achieve the above and other related objectives, the present invention also provides a risk entropy adaptive adjustment perception degradation safety planning system for implementing the aforementioned risk entropy adaptive adjustment perception degradation safety planning method. The system includes: a multi-source perception input terminal comprising: 1) an onboard environment perception module for receiving and processing a list of perceived targets and their corresponding perception confidence levels from cameras, millimeter-wave radar, and lidar; 2) a trajectory prediction module for outputting the multimodal predicted trajectory of each perceived target in the future time domain and its covariance matrix; and 3) a vehicle state estimation module for acquiring the vehicle's current position, speed, acceleration, heading angle, and yaw rate. The perception quality assessment module includes: 1) a sensor-level confidence fusion unit, which performs time-series smoothing on the category confidence of each perceived target to obtain the fused target-level perception confidence; 2) a scene-level perception quality scoring unit, which calculates the scene-level perception quality score by combining the confidence distribution of all perceived targets, the rate of change of the number of targets, and the detection missing rate; and 3) a perception degradation judgment unit, which compares the scene-level perception quality score with a preset degradation threshold to determine whether the current perception state is normal, degraded, or severely degraded. The risk entropy calculation engine includes: 1) an uncertainty propagation modeling unit, used to jointly model the confidence of the perception detection box and the covariance matrix of the predicted trajectory into a probability distribution; 2) a risk entropy calculation unit, used to calculate the risk entropy value describing the uncertainty of the vehicle's position based on the probability distribution; and 3) an entropy-confidence mapping unit, used to establish a quantitative mapping relationship between the risk entropy value and the perception confidence. The adaptive safe corridor generator includes: 1) a drivable area boundary extraction unit, used to extract the boundary constraints of the current drivable area from the high-precision map and perception results; 2) an expansion coefficient calculation unit, used to calculate the expansion coefficient of the safe corridor in real time based on the risk entropy value; and 3) a dynamic safe corridor construction unit, used to apply the expansion coefficient to the drivable area boundary to generate dynamic safe corridor constraints. The planning strategy hierarchical controller includes: 1) a planning strategy selection unit, which selects the current planning strategy from normal mode, early warning mode and conservative mode based on the perception degradation judgment result; 2) a mode switching smoothing unit, which performs weighted transition between adjacent planning strategies to avoid sudden changes in planning behavior; and 3) a strategy parameter configuration unit, which dynamically configures the parameter set of the constraint optimization planner according to the selected planning strategy. The constrained optimization planner includes: 1) a multi-objective cost function construction unit, used to construct a comprehensive cost function that integrates safety, smoothness, traffic efficiency and risk margin; 2) a hierarchical constraint solving unit, used to sort safety constraints, traffic rule constraints, dynamic constraints and comfort constraints by priority and solve them step by step; 3) a trajectory output unit, used to output the optimal trajectory that satisfies all priority constraints. The trajectory execution feedback unit includes: 1) a trajectory execution status monitoring unit, which is used to collect the tracking deviation between the actual trajectory of the vehicle and the planned trajectory in real time; 2) a deviation feedback correction unit, which is used to feed back the tracking deviation to the adaptive safety corridor generator and dynamically adjust the expansion coefficient; 3) a trigger-based replanning unit, which is used to start the local replanning process when the tracking deviation exceeds the preset threshold.

[0018] The present invention has the following positive effects: 1. This invention establishes a quantitative mapping relationship between perception confidence and risk entropy, quantifies perception uncertainty into risk entropy parameters that can participate in optimization calculations, realizes refined mathematical modeling of perception performance degradation, and solves the problem that perception quality information cannot enter the planning optimization solver in the prior art.

[0019] 2. This invention achieves a continuous dynamic correlation between safety margin and perceived quality by using an adaptive adjustment mechanism driven by risk entropy to adjust the expansion coefficient of the safety corridor. When perceived confidence decreases, the safety corridor automatically expands, and when perceived quality recovers, the safety corridor smoothly contracts, thus avoiding abrupt changes in planning behavior caused by the switching of discrete trajectory sets.

[0020] 3. This invention establishes a mapping mechanism between the perception degradation level and the planning strategy parameters through a hierarchical controller for planning strategies. When perception degradation occurs, it smoothly switches to the early warning mode or the conservative mode, maintaining traffic efficiency to the maximum extent while ensuring safety, thus achieving an adaptive balance between safety and traffic efficiency.

[0021] 4. This invention introduces an uncertainty cost term based on risk entropy into the trajectory optimization cost function, enabling the optimization solver to consider the impact of perceived uncertainty on safety constraints during the planning stage, rather than only performing trajectory selection after planning. This fundamentally improves the robustness of the planning method in scenarios of perceived degradation.

[0022] 5. This invention feeds back tracking deviations to the safety corridor generator in real time through a trajectory execution feedback unit, forming a complete closed loop from perception input to risk assessment, from optimization planning to execution feedback. When tracking performance deteriorates, the safety margin is automatically tightened, ensuring planning-execution consistency. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This diagram illustrates the working principle of the adaptive safety corridor dynamic generation mechanism of the present invention. Figure 3 This is a diagram illustrating the state switching logic of the hierarchical controller for the planning strategy of the present invention. Figure 4 This diagram illustrates the structure of the autonomous driving perception degradation safety planning system based on risk entropy adaptive adjustment according to the present invention. Figure 5 This diagram illustrates the internal structure and data flow of the perceived quality assessment module and risk entropy calculation engine of the present invention. Detailed Implementation

[0024] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0025] Example: Figure 1 As shown, a risk entropy adaptive adjustment-based perceived degradation safety planning method is proposed, the method comprising: M1. Reception and preprocessing of multi-source sensing data; M2. Perception quality assessment and perception degradation determination: Based on the perception confidence and covariance matrix output in step 1, three levels of assessment processing are performed. M3. Risk Entropy Calculation and Confidence-Inflation Coefficient Mapping: The risk entropy calculation engine jointly models the perceived confidence and the predicted covariance matrix as a probability distribution, calculates the risk entropy value H, and establishes a quantitative mapping relationship between risk entropy and the safety corridor inflation coefficient. M4. Adaptive safety corridor dynamic generation: Based on the safety corridor expansion coefficient output in step 3, and combined with the drivable area boundary constraints extracted from the high-precision map, a safety corridor is dynamically generated. M5. Planning strategy hierarchical selection and parameter configuration: Based on the perception degradation level output in step 2, select the current planning strategy from normal mode, early warning mode and conservative mode, and dynamically configure the parameter set of the constraint optimization planner. M6. Constraint optimization trajectory solution: Receive the adaptive safety corridor constraints generated in step 4 and the planning strategy parameters configured in step 5, construct and solve the multi-objective constrained optimization problem; M7. Trajectory Output and Execution Feedback: After the optimization solution converges, the constraint optimization planner outputs the optimal trajectory that satisfies all priority constraints, including the trajectory position sequence, the corresponding velocity sequence, and the acceleration sequence. M8. Perception Degradation Continuous Monitoring and Dynamic Update: Throughout the entire trajectory execution process, the perception quality assessment module continuously executes the assessment process of step 2 to monitor perception quality in real time.

[0026] In this embodiment, in step M1, the multi-source sensing data reception and preprocessing involves the input end receiving in real time a target list and corresponding sensing confidence scores from the vehicle-mounted sensing system; dynamic targets such as vehicles and pedestrians in front and to the side sensed by the camera, accompanied by target category confidence scores; target distance, relative speed, and angle information provided by millimeter-wave radar, accompanied by radial speed confidence scores; and high-precision three-dimensional sensing target point clouds generated by lidar, accompanied by point cloud density confidence scores. Simultaneously, the trajectory prediction module generates a multimodal predicted trajectory and its covariance matrix for each sensing target within a future set time domain.

[0027] In this embodiment, in step M2, the three-level evaluation process based on the perceived confidence and covariance matrix output in step 1 includes: M21. Sensor-level confidence fusion: For each perceived target i, the sensor-level confidence fusion unit performs exponentially weighted smoothing on the perceived confidence of its historical T frames. , Where α∈(0,1) are the fusion weight coefficients, c i (t) represents the original perceived confidence of the i-th target at the current time, and cfused i The confidence level of the target-level perception after fusion; M22. Scene-level perception quality score: The scene-level perception quality score unit integrates the fusion confidence distribution of all perceived targets, the rate of change in the number of targets, and the detection missing rate to calculate the scene-level perception quality score Q. , , in, r is the mean confidence score of all perceived targets fused together. miss r represents the detection missing rate of the current frame compared to the previous frame. drift The target quantity change rate is represented by w1, w2, and w3, which are weighting coefficients. When Q ≥ Q normal When Q is determined to be in a normal perception state, degraded ≤Q normal When Q is determined to be in a downgraded perception state, degraded The system is judged to be in a severely degraded perception state. degraded and Q normal The preset perceived quality threshold; M23. Perception Degradation Judgment and Level Output: The perception degradation judgment unit outputs the perception degradation level L based on the comparison between the scene-level perception quality score Q and the threshold. .

[0028] In this embodiment, the constraint condition for the weight coefficients w1, w2, and w3 is w1 + w2 + w3 = 1.

[0029] In this embodiment, step M3, the calculation of risk entropy and the mapping of confidence level to inflation coefficient, includes: M31. Uncertainty joint probability distribution modeling, fusing the confidence level c of each perceived target i. fused i And the predicted covariance matrix ∑ i Joint modeling is performed on the vehicle at each discrete position p. k The probability distribution p at point k (p), , Where, μ k The center position of the k-th discrete risk interval; M32. Risk entropy calculation, based on the probability distribution p in step 31. k (p), calculate the risk entropy value H, which describes the degree of uncertainty in the vehicle's location. , ​​The risk entropy H(p) ranges from [0, logK], where H(p) = 0 indicates that the perception is completely certain, and H(p) = logK indicates that the perception is completely uncertain. M33. Entropy-Confidence Mapping and Inflation Coefficient Calculation: Establishing a quantitative mapping relationship between the risk entropy value H(p) and the safety corridor inflation coefficient δ. When the perceived confidence decreases, the risk entropy increases, the safety corridor inflation coefficient increases accordingly, and the safety margin automatically expands. , Where, δ min λ is the minimum width of the safety corridor. H δ is the risk entropy amplification factor. k Given the current expansion coefficient of the safety corridor, δ increases as H(p) increases. k It increases linearly thereafter; when H(p) = 0, δ k =δ min It degenerates to the basic width under normal perception conditions.

[0030] In this embodiment, step M33 also introduces a correction for the expansion coefficient calculation based on the perception degradation level L. , Where, β L This is a correction factor for perception degradation level, used to further increase the safety margin when a severe perception degradation is detected.

[0031] In this embodiment, as Figure 2 As shown, in step M4, the drivable area boundary is used as the inner constraint, and the expanded risk area boundary is used as the outer constraint. The sequence convexization method is used to convexize the non-convex safety area into a series of convex polygon segments: , Among them, C safe,k For the k-th safe corridor, A k and b k Let p be the constraint matrix and constraint vector of the convex polygon, respectively. ref,k δ is the reference path point for this section of the corridor. k Let C be the expansion coefficient of this corridor segment, and C be the safety corridor constraint. safe It will serve as the hard constraint boundary for the constrained optimization planner.

[0032] In this embodiment, as Figure 3 As shown, in step M5, when the perception degradation level L=0, the normal mode is selected: all planning parameters remain at their default values, and the safety margin is δ. min Based on this, the traffic efficiency weight w e Maintain a high level; when L=1, select the early warning mode: increase the safety margin to δ.min +λ H • H(p), traffic efficiency weight w e Reduce the safety constraint weight w s Increase, while introducing an additional uncertainty penalty term w u • H(p) is incorporated into the cost function; when L=2, a conservative mode is selected: the safety margin is further increased to δ. min +λ H ∙H(p)+β L • L, traffic efficiency weight w e Reduced to the minimum, safety constraint weight w s Increase to the highest setting, limiting the maximum speed to v. safe max And increase the collision detection safety distance d safe .

[0033] In this embodiment, in step M6, the cost function J for the trajectory optimization problem is constructed by comprehensively considering four dimensions: safety, smoothness, traffic efficiency, and risk margin. , Among them, w u For uncertain penalty weights, J uncertainty For uncertainty cost terms based on risk entropy, , Security cost function J safety (τ) Penalty trajectory for crossing high-risk areas, smoothing cost function J smooth (τ) The jerk of the penalty trajectory is used to ensure ride comfort; the cost function J for traffic efficiency. efficiency (τ) Encourages the trajectory to reach the target point quickly, risk margin cost function J risk,margin (τ) The distance between the penalty trajectory and the high-risk area is insufficient.

[0034] In this embodiment, after the optimization solution converges, the constrained optimization planner outputs the optimal trajectory that satisfies all priority constraints, including the trajectory position sequence {p0, p1, ..., p...}. T The corresponding velocity sequence {v0, v1, ..., v} T} and acceleration sequence {a0,a1,...,a T The trajectory is sent to the vehicle's underlying controller for execution in the form of a target velocity vector or acceleration command.

[0035] The trajectory execution status monitoring module collects the vehicle's actual driving trajectory in real time and calculates its tracking deviation from the planned trajectory: , When the tracking deviation exceeds a preset threshold, the deviation information is fed back to the adaptive safety corridor generator, and the safety corridor expansion coefficient is dynamically adjusted according to the following rules: , Where η is the deviation feedback gain coefficient. When the deviation continues to increase, the safety corridor automatically shrinks to increase the tracking performance incentive; when the deviation falls back, the safety corridor gradually recovers to the target expansion level driven by risk entropy. At the same time, when the tracking deviation exceeds the replanning trigger threshold, the trigger-based replanning unit starts the local replanning process and re-executes the optimization solution process from steps M4 to M7.

[0036] In this embodiment, throughout the trajectory execution process, the perception quality assessment module continuously executes the assessment process of step M2 to monitor perception quality in real time. When the perception degradation level L changes, the level change information is simultaneously transmitted to the risk entropy calculation engine and the planning strategy hierarchical controller, triggering real-time recalculation of the safety corridor expansion coefficient and dynamic switching of planning strategy parameters. When the perception quality recovers (L changes from a high level to a low level or Q rises back to Q), the perception quality improves. normal When the above conditions are met, the system will automatically and smoothly switch back to normal mode without manual intervention.

[0037] In this embodiment, as Figure 4 As shown, the present invention also provides a risk entropy adaptive adjustment perception degradation safety planning system for implementing the risk entropy adaptive adjustment perception degradation safety planning method. The system includes: a multi-source perception input terminal comprising: 1) an on-board environment perception module for receiving and processing a list of perceived targets and their corresponding perception confidence levels from cameras, millimeter-wave radar, and lidar; 2) a trajectory prediction module for outputting the multimodal predicted trajectory of each perceived target in the future time domain and its covariance matrix; and 3) a vehicle state estimation module for obtaining the vehicle's current position, speed, acceleration, heading angle, and yaw rate. The perception quality assessment module includes: 1) a sensor-level confidence fusion unit, which performs time-series smoothing on the category confidence of each perceived target to obtain the fused target-level perception confidence; 2) a scene-level perception quality scoring unit, which calculates the scene-level perception quality score by combining the confidence distribution of all perceived targets, the rate of change of the number of targets, and the detection missing rate; and 3) a perception degradation judgment unit, which compares the scene-level perception quality score with a preset degradation threshold to determine whether the current perception state is normal, degraded, or severely degraded. In this embodiment, as Figure 5As shown, the risk entropy calculation engine includes: 1) an uncertainty propagation modeling unit, used to jointly model the confidence of the perception detection box and the covariance matrix of the predicted trajectory into a probability distribution; 2) a risk entropy calculation unit, used to calculate the risk entropy value describing the uncertainty of the vehicle's position based on the probability distribution; and 3) an entropy-confidence mapping unit, used to establish a quantitative mapping relationship between the risk entropy value and the perception confidence. The adaptive safe corridor generator includes: 1) a drivable area boundary extraction unit, used to extract the boundary constraints of the current drivable area from the high-precision map and perception results; 2) an expansion coefficient calculation unit, used to calculate the expansion coefficient of the safe corridor in real time based on the risk entropy value; and 3) a dynamic safe corridor construction unit, used to apply the expansion coefficient to the drivable area boundary to generate dynamic safe corridor constraints. The planning strategy hierarchical controller includes: 1) a planning strategy selection unit, which selects the current planning strategy from normal mode, early warning mode and conservative mode based on the perception degradation judgment result; 2) a mode switching smoothing unit, which performs weighted transition between adjacent planning strategies to avoid sudden changes in planning behavior; and 3) a strategy parameter configuration unit, which dynamically configures the parameter set of the constraint optimization planner according to the selected planning strategy. The constrained optimization planner includes: 1) a multi-objective cost function construction unit, used to construct a comprehensive cost function that integrates safety, smoothness, traffic efficiency and risk margin; 2) a hierarchical constraint solving unit, used to sort safety constraints, traffic rule constraints, dynamic constraints and comfort constraints by priority and solve them step by step; 3) a trajectory output unit, used to output the optimal trajectory that satisfies all priority constraints. The trajectory execution feedback unit includes: 1) a trajectory execution status monitoring unit, which is used to collect the tracking deviation between the actual trajectory of the vehicle and the planned trajectory in real time; 2) a deviation feedback correction unit, which is used to feed back the tracking deviation to the adaptive safety corridor generator and dynamically adjust the expansion coefficient; 3) a trigger-based replanning unit, which is used to start the local replanning process when the tracking deviation exceeds the preset threshold.

[0038] In summary, this invention not only solves the problem that perceived quality information cannot enter the planning optimization solver in the prior art, but also avoids abrupt changes in planning behavior caused by the switching of discrete trajectory sets.

[0039] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A risk entropy adaptive adjustment-based perception degradation safety planning method, characterized in that, The method includes: M1. Reception and preprocessing of multi-source sensing data; M2. Perception quality assessment and perception degradation determination: Based on the perception confidence and covariance matrix output in step 1, three levels of assessment processing are performed. M3. Risk Entropy Calculation and Confidence-Inflation Coefficient Mapping: The risk entropy calculation engine jointly models the perceived confidence and the predicted covariance matrix as a probability distribution, calculates the risk entropy value H, and establishes a quantitative mapping relationship between risk entropy and the safety corridor inflation coefficient. M4. Adaptive safety corridor dynamic generation: Based on the safety corridor expansion coefficient output in step 3, and combined with the drivable area boundary constraints extracted from the high-precision map, a safety corridor is dynamically generated. M5. Planning strategy hierarchical selection and parameter configuration: Based on the perception degradation level output in step 2, select the current planning strategy from normal mode, early warning mode and conservative mode, and dynamically configure the parameter set of the constraint optimization planner. M6. Constraint optimization trajectory solution: Receive the adaptive safety corridor constraints generated in step 4 and the planning strategy parameters configured in step 5, construct and solve the multi-objective constrained optimization problem; M7. Trajectory Output and Execution Feedback: After the optimization solution converges, the constraint optimization planner outputs the optimal trajectory that satisfies all priority constraints, including the trajectory position sequence, the corresponding velocity sequence, and the acceleration sequence. M8. Perception Degradation Continuous Monitoring and Dynamic Update: Throughout the entire trajectory execution process, the perception quality assessment module continuously executes the assessment process of step 2 to monitor perception quality in real time.

2. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M1, the multi-source sensing data reception and preprocessing involves the input end receiving a target list and corresponding sensing confidence scores from the vehicle-mounted sensing system in real time, and the dynamic targets of vehicles and pedestrians in front and to the side sensed by the camera, along with target category confidence scores. Millimeter-wave radar provides target distance, relative velocity, and angle information along with radial velocity confidence; lidar generates high-precision three-dimensional perception target point clouds along with point cloud density confidence; and the trajectory prediction module generates multimodal predicted trajectories and their covariance matrices for each perceived target within a future set time domain.

3. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M2, the three-level evaluation process based on the perceived confidence and covariance matrix output in step 1 includes: M21. Sensor-level confidence fusion: For each perceived target i, the sensor-level confidence fusion unit performs exponentially weighted smoothing on the perceived confidence of its historical T frames. ， Where α∈(0,1) are the fusion weight coefficients, c i (t) represents the original perceived confidence of the i-th target at the current time, and c fused i The confidence level of the target-level perception after fusion; M22. Scene-level perception quality score: The scene-level perception quality score unit integrates the fusion confidence distribution of all perceived targets, the rate of change in the number of targets, and the detection missing rate to calculate the scene-level perception quality score Q. ， ， in, r is the mean confidence score of all perceived targets fused together. miss r represents the detection missing rate of the current frame compared to the previous frame. drift The target quantity change rate is represented by w1, w2, and w3, which are weighting coefficients. When Q ≥ Q normal When Q is determined to be in a normal perception state, degraded ≤Q normal When Q is determined to be in a downgraded perception state, degraded The system is judged to be in a severely degraded perception state. degraded and Q normal The preset perceived quality threshold;​​ M23. Perception Degradation Judgment and Level Output: The perception degradation judgment unit outputs the perception degradation level L based on the comparison between the scene-level perception quality score Q and the threshold. 。 4. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 3, characterized in that: The constraint condition for the weight coefficients w1, w2, and w3 is w1 + w2 + w3 = 1.

5. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M3, the risk entropy calculation and confidence-inflation coefficient mapping includes: M31. Uncertainty joint probability distribution modeling, fusing the confidence level c of each perceived target i. fused i And the predicted covariance matrix ∑ i Joint modeling is performed on the vehicle at each discrete position p. k The probability distribution p at point k (p), ， Where, μ k The center position of the k-th discrete risk interval; M32. Risk entropy calculation, based on the probability distribution p in step 31. k (p), calculate the risk entropy value H, which describes the degree of uncertainty in the vehicle's location. ， The risk entropy H(p) ranges from [0, logK], where H(p) = 0 indicates that the perception is completely certain, and H(p) = logK indicates that the perception is completely uncertain. M33. Entropy-Confidence Mapping and Inflation Coefficient Calculation: Establishing a quantitative mapping relationship between the risk entropy value H(p) and the safety corridor inflation coefficient δ. When the perceived confidence decreases, the risk entropy increases, the safety corridor inflation coefficient increases accordingly, and the safety margin automatically expands. ， Where, δ min λ is the minimum width of the safety corridor. H δ is the risk entropy amplification factor. k Given the current expansion coefficient of the safety corridor, δ increases as H(p) increases. k It increases linearly thereafter; when H(p) = 0, δ k =δ min It degenerates to the basic width under normal perception conditions.

6. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 5, characterized in that: In step M33, a correction to the expansion coefficient calculation is also introduced based on the perception degradation level L. ， Where, β L This is a correction factor for perception degradation level, used to further increase the safety margin when a severe perception degradation is detected.

7. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M4, using the drivable area boundary as the inner constraint and the expanded risk area boundary as the outer constraint, the non-convex safety area is convexified into a series of convex polygon segments using a sequential convexification method: ， Among them, C safe,k For the k-th safe corridor, A k and b k Let p be the constraint matrix and constraint vector of the convex polygon, respectively. ref,k δ is the reference path point for this section of the corridor. k Let C be the expansion coefficient of this corridor segment, and C be the safety corridor constraint. safe It will serve as the hard constraint boundary for the constrained optimization planner.

8. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M5, when the perception degradation level L=0, the normal mode is selected: all planning parameters remain at their default values, and the safety margin is δ. min Based on this, the traffic efficiency weight w e Maintain a high level; When L=1, the early warning mode is selected: the safety margin is increased to δ. min +λ H • H(p), traffic efficiency weight w e Reduce the safety constraint weight w s Increase, while introducing an additional uncertainty penalty term w u • H(p) is incorporated into the cost function; when L=2, a conservative mode is selected: the safety margin is further increased to δ. min +λ H ∙H(p)+β L • L, traffic efficiency weight w e Reduced to the minimum, safety constraint weight w s Increase to the highest setting, limiting the maximum speed to v. safe max And increase the collision detection safety distance d safe .

9. The risk entropy adaptive adjustment perception degradation safety planning method according to claim 1, characterized in that, In step M6, considering four dimensions—safety, smoothness, traffic efficiency, and risk margin—the cost function J for the trajectory optimization problem is constructed. ， Among them, w u For uncertain penalty weights, J uncertainty For uncertainty cost terms based on risk entropy, ， Security cost function J safety (τ) Penalty trajectory for crossing high-risk areas, smoothing cost function J smooth (τ) The jerk of the penalty trajectory is used to ensure ride comfort; the cost function J for traffic efficiency. efficiency (τ) Encourages the trajectory to reach the target point quickly, risk margin cost function J risk,margin (τ) The distance between the penalty trajectory and the high-risk area is insufficient.

10. A risk entropy adaptive adjustment perception degradation safety planning system, characterized in that, The system is used to implement the risk entropy adaptive adjustment perception degradation safety planning method according to any one of claims 1-9. The system includes: a multi-source perception input terminal comprising: 1) an on-board environment perception module, used to receive and process a list of perceived targets and corresponding perception confidence levels from a camera, millimeter-wave radar, and lidar; 2) a trajectory prediction module, used to output the multimodal predicted trajectory of each perceived target in the future time domain and its covariance matrix; 3) a vehicle state estimation module, used to obtain the current position, speed, acceleration, heading angle, and yaw rate of the vehicle. The perception quality assessment module includes: 1) a sensor-level confidence fusion unit, which performs time-series smoothing on the category confidence of each perceived target to obtain the fused target-level perception confidence; 2) a scene-level perception quality scoring unit, which calculates the scene-level perception quality score by combining the confidence distribution of all perceived targets, the rate of change of the number of targets, and the detection missing rate; and 3) a perception degradation judgment unit, which compares the scene-level perception quality score with a preset degradation threshold to determine whether the current perception state is normal, degraded, or severely degraded. The risk entropy calculation engine includes: 1) an uncertainty propagation modeling unit, used to jointly model the confidence of the perception detection box and the covariance matrix of the predicted trajectory into a probability distribution; 2) a risk entropy calculation unit, used to calculate the risk entropy value describing the uncertainty of the vehicle's position based on the probability distribution; and 3) an entropy-confidence mapping unit, used to establish a quantitative mapping relationship between the risk entropy value and the perception confidence. The adaptive safe corridor generator includes: 1) a drivable area boundary extraction unit, used to extract the boundary constraints of the current drivable area from the high-precision map and perception results; 2) an expansion coefficient calculation unit, used to calculate the expansion coefficient of the safe corridor in real time based on the risk entropy value; and 3) a dynamic safe corridor construction unit, used to apply the expansion coefficient to the drivable area boundary to generate dynamic safe corridor constraints. The planning strategy hierarchical controller includes: 1) a planning strategy selection unit, which selects the current planning strategy from normal mode, early warning mode and conservative mode based on the perception degradation judgment result; 2) a mode switching smoothing unit, which performs weighted transition between adjacent planning strategies to avoid sudden changes in planning behavior; and 3) a strategy parameter configuration unit, which dynamically configures the parameter set of the constraint optimization planner according to the selected planning strategy. The constrained optimization planner includes: 1) a multi-objective cost function construction unit, used to construct a comprehensive cost function that integrates safety, smoothness, traffic efficiency and risk margin; 2) a hierarchical constraint solving unit, used to sort safety constraints, traffic rule constraints, dynamic constraints and comfort constraints by priority and solve them step by step; 3) a trajectory output unit, used to output the optimal trajectory that satisfies all priority constraints. The trajectory execution feedback unit includes: 1) a trajectory execution status monitoring unit, which is used to collect the tracking deviation between the actual trajectory of the vehicle and the planned trajectory in real time; 2) a deviation feedback correction unit, which is used to feed back the tracking deviation to the adaptive safety corridor generator and dynamically adjust the expansion coefficient; 3) a trigger-based replanning unit, which is used to start the local replanning process when the tracking deviation exceeds the preset threshold.

Images